Exploring LoRA Training: Tuning Parameters for 2400 Steps with 40 Images

Dark_Sytze · reddit · 2026-07-24

A developer shared practical experience in training LoRA for consistent game character art using Anima Trainflow. The dataset contains 40-50 character images, with settings including Network rank 32, Batch size 1, and Training steps 2400.

Using the Prodigy optimizer, the author stopped the training early at 1200 steps due to long processing times. The post primarily discusses the specific impact of Gradient accumulation on VRAM usage and training time, seeking community advice on parameter tuning.

Original post →

More from Multimodal

Multimodal channel →